AI Connect · Rīga · 22.08.2026 · Lightning talk · 9:00
How we learned to control AI hallucinations in data analytics — by running an AI data analyst in production.
Albinas Plešnys — co-founder, tvaras.ai · Matas Baltrėnas — Google Ads Impact Awards 2025 · AI Excellence
A quick show of hands
Keep them up. This talk is about why that happens — and what fixes it.
The economics of a wrong number
Nobody needs just numbers. Companies don't pay for charts — they pay for truth in them. The moment one number is invented, every number needs re-checking, and the value drops to zero.
The biggest failure: AI doesn't know
It doesn't know things — it generates plausible text. Often the plausible happens to be true. In analytics, "often" is exactly the problem.
A short philosophical detour — knowledge = justified true belief
Can an LLM hold beliefs? We don't know — and it doesn't matter. What matters is not that the LLM knows. It's that we do.
To hold a belief you first need it formulated. The LLM states a number.
← the LLMThe LLM must show how it got there — definitions used, queries run, sources read. All of it.
← the LLMWe verify the claim ourselves — using the track the LLM provided. The human closes the loop.
← the humanGettier cases graciously ignored — we have 9 minutes.
Where LLM justification breaks down
Some definitions are genuinely ambiguous — the model picks one silently and justifies the wrong question."revenue" → gross? net? booked? recognised? incl. refunds?
Some answers depend on the business, not the data. "Which campaigns underperform?" depends on your target ROAS — which depends on how much of revenue is recurring.same query · two companies · opposite verdicts
Some data is genuinely missing — and the model "helpfully" invents a number rather than admitting the void.empty result → "≈ €48,200"
The exact metric doesn't exist — so the model silently substitutes a proxy and answers as if it were the real thing.asked: LTV → answered: avg. order value × a guess
What didn't work
More context, more instructions — the model still guessed, just more verbosely.
Telling a model to not hallucinate is a hope, not a mechanism.
Hand-written context files drift from reality the day after they're written. Rules must come from the live data stack, not from documents about it.
Verification hidden inside the machine produces confidence, not knowledge. If no human can check the track, nobody knows anything.
What actually worked
Built automatically from the data model + exposures — your existing dashboards already "know" the truth. The AI joins the data/BI stack that holds the business rules; it doesn't start from zero.
Learning from mistakes and successes. Corrections and confirmations accumulate — the same question doesn't get re-litigated every Monday.
SQL + reasoning behind every number, exposed for humans to check. Each confirmation strengthens the memories it came from.
Put together
Not completely automated — deliberately. But 10× faster than doing analytics alone. That is how AI wins in data.
In action — every number carries its justification
Numbers are claims. Claims carry their track. R replays the demo.
Try to make it invent a number — demo zone, today
Albinas Plešnys · Matas Baltrėnas · hello@vej.ai